Sign up for our newsletter
AI

Tomorrow’s bread: How Hovis predicts demand for a million loaves a day

The UK breadmaker's forecasting algorithm predicts demand for every customer address for up to a year, every 24 hours.

Every business needs to match its output to customer demand but it is especially vital in baking: nobody wants last week’s bread. Hovis, the UK’s iconic bakery brand, delivers more than a million loaves every day, so even slight misjudgements in output can lead to considerable waste.

Tech Monitor spoke to Harry Watts, head of sales and operations planning and data science about how data analytics at Hovis has evolved to the point that it can predict demand in individual postcodes for up to a year in advance – every 24 hours.

How Hovis predicts bread demand
Hovis’s Icarus platform allows it to reforecast its business for up to a year, every day. (Photo by Phil Benford/Shutterstock)

The bedrock of Hovis’s analytics capabilities is extensive data collection. “We capture every minute of each of our bakeries,” explains Watts. “We count the number of loaves going through the plant and capture all of our customer orders from past years and all the individual addresses for those orders.”

In 2018, Watts and his team adopted data visualisation software from Tableau to analyse this data. Tableau was introduced to address the challenges of processing millions of rows of data in Excel spreadsheets. “We needed a proper business intelligence solution to help us with that,” says Watts.

White papers from our partners

Tableau allows the team to analyse live data, as well as historic reports. The system draws from the company’s Azure-hosted data warehouse, which means that as soon as an order is received, it is available for analysis. “This has been a game-changing component of the project as it allows us to understand where we are in terms of our orders, our data and everything that we’ve got,” Watts says.

This up-to-the-minute reporting allows Watts’ team to spot anomalies that might predict customer service issues. “We are now able to talk to customers at very short notice from the order being placed as sometimes customers place orders which look a bit strange or unexpected – maybe they pressed the button twice and ordered more product than they needed. The last thing we want is customers not getting their orders or getting the wrong products. With Tableau, we are able to quickly validate those orders.”

This Tableau-based system allowed Hovis’s demand-planning function to achieve 94% forecasting accuracy – “the best I’ve seen in my history,” says Watts.

Granular predictions

However, the forecasts this system produced were aggregated, which is fine for high-level commercial planning but unable to predict demand for individual customers or locations. But Watts knew there was more value to be extracted from the company’s data. “We have the data and a good way to visualise it,” Watts recalls thinking. “How do we leverage that data to enrich it and build greater insights?”

“This is where we landed ourselves within Databricks,” he says.

Databricks is a cloud platform which allows data scientists and data engineers to collaborate and experiment with analytics projects. The amalgamation of Tableau and Databricks gave birth to Icarus – Hovis’s new machine learning-powered forecasting platform, which was launched in September last year.

Icarus has allowed Watts’ team to develop a forecasting algorithm that draws on data about individual orders, their component products and destination addresses – a far greater level of granularity than it could previously handle.

“Icarus can analyse tens of thousands of combinations of individual products and addresses forecasts,” explains Watts. “We can return a forecast for every address/product combination in our estate that will provide us six to 12 months’ view of what the likely demand is at that address level. That’s been immensely powerful.”

Icarus allows the team to introduce new data series, such as weather data, and assess whether they are correlated with demand for a particular address. It can incorporate factors such as promotions, holidays or lockdown restrictions and other Covid-19-related factors that may shape demand. “We can identify each of those regresses and ultimately, Icarus lets you validate and build the forecasting components,” says Watts.

The predictions that Icarus produces integrate neatly into Hovis’s existing production schedule, Watts says. “We haven’t fundamentally changed our logistics and our manufacturing plans,” he explains. “We just come in at the right moment with the right numbers, and I think that’s really what’s been quite transformational about this.”

As a result of Icarus, Hovis can now reforecast its business for up to a year, every single day, Watts says. “We will create a new forecast for each of the addresses in our database running out for six to 12 months,” he says. “We refocus the entire business every day.”

Cristina Lago

Associate editor

Cristina Lago is associate editor of Tech Monitor.